CN101493886B - Karyoplast categorization and identification method in case of unsoundness of characteristic parameter - Google Patents

Abstract

The invention relates to a caryon classification and recognition method with incomplete characteristic parameters. The method comprises the following steps: a. in a set of caryon training samples, statistics is adopted to identify whether the characteristic parameters are volatile or nonvolatile; b. during training, in the set of caryon training samples, a single volatile characteristic value is taken as output and the nonvolatile characteristics as input so as to obtain a regression estimate mode of the single volatile characteristic value by supporting vector regression SVR training program; c. during recognition, as to a single caryon, if a missing characteristic parameter is available, regression estimate is conducted to the missing characteristic parameter by utilizing the regressionestimate mode obtained from training and a complete characteristic parameter of the regression estimate mode; and d. the missing characteristic parameter after filling is combined with the complete characteristic parameter to form a new caryon characteristic vector which is used for conducting the recognition of caryon classification.

Description

Nucleus classification and recognition methods under the imperfect situation of characteristic parameter

Technical field

The present invention relates to nucleus classification and recognition methods under the imperfect situation of a kind of characteristic parameter, belong to the biomedical engineering technology field.

Background technology

Pair cell nuclear carries out the core that Classification and Identification is a cell quantitative, also is an indispensable ring in automatic examination of quantitative cytology cancer and the pathology automatic diagnosis process simultaneously.Utilizing characteristics such as nuclear optics, texture, form, color, relation, through the type of SVMs, neural network or k nearest neighbor algorithm isotype recognition methods recognizing cells nuclear, is the topmost method of nucleus Classification and Identification.Fig. 1 provides the main flow process of nucleus Classification and Identification.Can see that from Fig. 1 nuclear Classification and Identification is divided into two independently processes: training process and identifying.In the system research and development process; Utilization is carried out the classification results that artificial cognition obtains through the characteristic parameter (a plurality of characteristic parameter composition characteristics vector) of nuclear image calculation acquisition with by the pathology expert pair cell nuclear of specialty; Training program through pattern-recognition obtains nuclear disaggregated model, and this process is called training process.In the process of the actual use of system; The nucleus disaggregated model that obtains in nucleus proper vector that same method calculates in utilization and the training process and the training process; Through the recognizer in the pattern-recognition; Can obtain nuclear automatic classification results, this process is called identifying.

From nuclear image, the characteristic parameter that can extract is a lot, and generally they can be divided into several big type: morphological feature, textural characteristics, optical signature, color characteristic and relationship characteristic etc.Through extracting these characteristic parameter composition characteristic vectors, for the nucleus Classification and Identification provides foundation.In general, the quantity of characteristic parameter is many more, and the effect of nucleus Classification and Identification is also good more, and the result is accurate more.

But; On the one hand; Because the problem of IMAQ instrument [such as the necrosis point in instrument failure, CCD (the Charge Coupled Device) picture pick-up device etc.], the perhaps restriction of the computing method of characteristic parameter own is not can both calculate all characteristic parameters for all nuclei pictures.In the parts of fine karyon, having the Partial Feature parameter can't obtain, and the also difference with nuclear difference of the characteristic parameter of these disappearances.On the other hand, for most algorithm for pattern recognition, generally all do not provide the training and the recognition methods of losing under the characteristic parameter situation.So just brought difficulty for the classification of utilizing algorithm for pattern recognition recognizing cells nuclear accurately.

In the Chinese patent 03131975 " the lung carcinoma cell image-recognizing method of a kind of high precision, low false negative rate "; This method utilizes the digital camera of settling on the optical microscope to take the cell pathology section; The vision signal that digital camera produces gets into computing machine behind image capture device; After through the appropriate image pre-service, give the lung carcinoma cell image recognition with the image of cell and partly handle, wherein mention and adopt the recognition methods of neural network as cancer; Be easy to generate the over-fitting phenomenon, cause the popularization of model of cognition limited in one's ability.In the structure (training process) of model of cognition, need artificially to participate in, workload is big.And do not provide the cell recognition method under the characteristic parameter deletion condition.And in the Chinese patent 200710192233 " Intelligentize lung cancer early cell pathological picture recognition processing method "; Provided the general frame of cell recognition and diagnosing tumor; Be divided into image pre-service, image segmentation, superpose cell separation reconstruct, feature extraction and selection and five steps of cytological classification to the pair cell treatment of picture, but do not provide definite description for the recognition methods and the process of cell.Particularly under the characteristic parameter deletion condition, how nucleus classifies and identification, does not provide corresponding method.

Summary of the invention

The object of the invention just provides nucleus classification and the recognition methods under the imperfect situation of a kind of characteristic parameter; Under the situation that it is imperfect at characteristic parameter, the Partial Feature parameter lacks; Estimate to fill the characteristic parameter of disappearance, complete and accurate carry out the nucleus Classification and Identification.

Technical scheme of the present invention is: nucleus classification and recognition methods under the imperfect situation of characteristic parameter; May further comprise the steps: a, concentrated at the nucleus training sample; Distinguishing which characteristic parameter through statistics is to be prone to lose characteristic, and which characteristic parameter is to be difficult for losing characteristic; It is characterized in that: also comprise in b, the training process; Concentrate at the nucleus training sample; Utilize single easy mistake characteristic ginseng value as output, be difficult for losing characteristic, obtain the recurrence estimation model of single easy mistake characteristic ginseng value through support vector regression SVR training program as input; In c, the identifying, for individual cells nuclear, if the disappearance characteristic parameter is arranged, recurrence estimation model that the utilization training obtains and its not disappearance characteristic parameter return and estimate to be somebody's turn to do the disappearance characteristic parameter; D, the disappearance characteristic parameter that utilizes process to fill in conjunction with the characteristic parameter that does not lack, are formed new nucleus proper vector, utilize this proper vector to carry out nuclear Classification and Identification.

Nucleus classification and recognition methods under the imperfect situation of characteristic parameter may further comprise the steps: a, concentrated at the nucleus training sample, and distinguishing which characteristic parameter through statistics is to be prone to lose characteristic, which characteristic parameter is the difficult characteristic of losing; It is characterized in that: also comprise

In b, the training process, concentrate, utilize single easy mistake characteristic ginseng value, be difficult for losing characteristic, obtain the recurrence estimation model of single easy mistake characteristic ginseng value through support vector regression SVR training program as input as output at the nucleus training sample;

In c, the identifying, for individual cells nuclear, if the disappearance characteristic parameter is arranged, recurrence estimation model that the utilization training obtains and its not disappearance characteristic parameter return and estimate to be somebody's turn to do the disappearance characteristic parameter;

Specifically: easy mistake characteristic

disappearance of supposition nucleus f calculates estimated value through following method and fills;

(1) extracts this nuclear difficult mistake proper vector b ⁱ

(2) the characteristic parameter a that utilizes training process to obtain _jRecurrence estimation model r _j() returns program through SVR, obtains the estimated value of this characteristic parameter

(3) using the estimated value

fill in the missing characteristic parameter values

Characteristic parameter for these all disappearances of nucleus; All utilize the method for above step (1), (2), (3) to obtain estimated value and fill; After having filled all disappearance characteristic parameters; In conjunction with the easy mistake characteristic parameter that does not lack, obtain being prone to lose the set

of all estimated values of proper vector

D, last; Utilize the difficult mistake proper vector of populated easy mistake proper vector and former extraction, the proper vector

of forming nucleus i jointly is used for nuclear Classification and Identification;

In the above-mentioned steps

I is used for distinguishing different cells nuclear, and j is used for distinguishing the different character parameter,

Represent that i nuclear j is prone to lose characteristic,

Represent i nuclear j estimated value that is prone to lose characteristic;

Represent i nuclear all easy set of losing the characteristic parameter estimated value, b ⁱRepresent i nuclear difficult mistake proper vector; a _jThe title of representing j characteristic parameter; f ⁱRepresent the set of i all characteristic parameters of nucleus.

Principle of work of the present invention is: utilize the known features parameter as input, the unknown characteristics parameter obtains the recurrence estimation model of unknown characteristics parameter as output through the SVR training program; Utilize the known features parameter as input, return estimation model, estimate the numerical value of unknown characteristics parameter and fill through SVR recurrence program as model; The characteristic parameter that utilizes populated disappearance characteristic parameter and do not lack is formed new proper vector, utilizes this proper vector to carry out nuclear Classification and Identification.

The invention has the beneficial effects as follows: (1) the present invention utilizes the known characteristic ginseng value of nucleus; The characteristic ginseng value that the method estimation that returns through SVR lacks; Effectively utilized the redundancy between the nucleus different characteristic parameter; But made full use of nuclear information extraction, improved the accuracy of identification; (2) the present invention has made full use of all utilizable resources; Under the situation that nuclear disappearance characteristic parameter is not effectively filled; Nuclear recognizer can not utilize the information of this characteristic parameter; Thereby the information in the nucleus do not lack this parameter that makes can not be fully used, and the present invention is fine has solved this problem; (3) the present invention utilizes SVR (Support Vector Regression) method to return the numerical value of disappearance characteristic parameter, has taken into full account linearity and nonlinear relationship between each characteristic parameter.

Description of drawings

Fig. 1 is the main process flow diagram of nucleus Classification and Identification.

Fig. 2 is the training process synoptic diagram of embodiment of the invention disappearance characteristic parameter regression model.

Fig. 3 is the filling process synoptic diagram of embodiment of the invention disappearance characteristic parameter.

Fig. 4 is the easy mistake characteristic parameter estimation procedure synoptic diagram in the embodiment of the invention 1.

Fig. 5 is the easy mistake characteristic parameter regression model training process synoptic diagram in the embodiment of the invention 1.

Embodiment

Below in conjunction with accompanying drawing and embodiment the present invention is done further explanation.

The present invention solves following problem:

(1) for training process a kind of suitable characteristic parameter fill method is provided; Under the situation of nuclear Partial Feature parameter disappearance; Can utilize suitable value to fill these parameters, so that the pattern-recognition training program can be carried out normal nucleus model of cognition training process;

(2), under the situation of nuclear Partial Feature parameter disappearance, can utilize suitable value to fill these parameters, so that the pattern recognition classifier program can be carried out normal nucleus identifying for identifying provides a kind of suitable characteristic parameter fill method;

(3) a kind of method is provided, can uses up the how possible quantity of utilizing characteristic parameter, carry out nuclear identification.

The practical implementation of the embodiment of the invention mainly comprises the content of the following aspects:

At first, find characteristic parameter (being prone to lose characteristic) that might lack and the characteristic parameter (being difficult for losing characteristic) that can not lack, set up the regression model of disappearance characteristic parameter.Detailed process such as Fig. 2.

For an extractible n characteristic parameter in the nucleus, supposing has l for being difficult for losing characteristic, and m for being prone to lose characteristic, m+l=n.Wherein, i the easy proper vector of losing of nucleus can be expressed as

Being difficult for losing proper vector does $b^i={(b_1^i,b_2^i,{\·\·\·b}_l^i)}^T.$

Be prone to lose characteristic for each, we can find N _jIndividual this characteristic does not have the nucleus of disappearance, forms the input and output collection

Wherein i is used for distinguishing different cells nuclear,

Represent that i nuclear j is prone to lose characteristic, as the output of regression model; b ⁱRepresent i nuclear difficult mistake proper vector, as the input of regression model.This input and output collection through a SVR regression model training aids, just can be obtained characteristic parameter a _jRegression model r _j().

Secondly, utilize regression model and nuclear difficult mistake characteristic ginseng value, return program, return and estimate the characteristic ginseng value that this nucleus lacks through SVR.Detailed process such as Fig. 3.

Suppose nuclear characteristic ginseng value

disappearance, can calculate estimated value through following method and fill.

(1) extracts this nuclear difficult mistake proper vector b ⁱ

(2) the characteristic parameter a that utilizes training process to obtain _jRegression model r _j() returns program through SVR, obtains the estimated value of this characteristic parameter

(3) utilization value is filled the characteristic ginseng value

of disappearance

For the characteristic parameter of these all disappearances of nucleus, all utilize the method for above step (1), (2), (3) to obtain estimated value and fill.After having filled all disappearance characteristic parameters; In conjunction with the easy mistake characteristic parameter that does not lack, obtain being prone to lose the set

of all estimated values of proper vector

At last; Can utilize the difficult mistake proper vector of populated easy mistake proper vector and former extraction, the proper vector

of forming nucleus i jointly is used for nuclear Classification and Identification.

Utilizing nuclear micro-image to carry out in the early diagnosis process of tumour,, may cause the excalation of cell image, thereby cause the disappearance of cell characteristic parameter owing to reasons such as the fault of CCD collecting device, interim inefficacies.When this situation occurs, can utilize the present invention to carry out the estimation of characteristic parameter, fill up the characteristic parameter of disappearance, thereby discern normally and diagnostic procedure.

In the quantitative cell analysis system of embodiment 1:DNA, the Partial Feature parameter is to be to obtain under some situation.For example high DNA content of material district mean radius, middle dna material content mean radius, low DNA content of material district mean radius:

$\mathrm{HighDNAAvgRadius}=\frac{\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}{R}_{x,y}{\mathrm{\Ω}}_{x,y}^{\mathrm{high}}}{\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}{\mathrm{\Ω}}_{x,y}^{\mathrm{high}}}$

$\mathrm{MedianDNAAvgRadius}=\frac{\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}{R}_{x,y}{\mathrm{\Ω}}_{x,y}^{\mathrm{median}}}{\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}{\mathrm{\Ω}}_{x,y}^{\mathrm{median}}}$

$\mathrm{LowDNAAvgRadius}=\frac{\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}{R}_{x,y}{\mathrm{\Ω}}_{x,y}^{\mathrm{low}}}{\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}{\mathrm{\Ω}}_{x,y}^{\mathrm{low}}}$

Wherein, R _{X, y}(x is y) with the distance at nucleus center for the expression pixel;

Be respectively the mask figure in high, medium and low DNA district, (x y) belongs to this zone, and this is estimated as 1, otherwise is 0 as if pixel.Can find out that if area that should the zone is 0 (very common in actual detected), mean radius result of calculation that then should the zone is illegal value (remove zero illegal), thereby cause the disappearance of this characteristic parameter.

Under this situation, can utilize the embodiment of the invention that this characteristic parameter is estimated, thus the characteristic parameter that obtains estimating.

HighDNAAvgRadius＝r _{HighDNAAvgRadius}(b)

MedianDNAAvgRadius＝r _{MedianDNAAvgRadius}(b)

LowDNAAvgRadius＝r _{LowDNAAvgRadius}(b)

Wherein, b is for being difficult for losing characteristic ginseng value.After obtaining corresponding estimated value; Can be difficult for to lose characteristic ginseng value; The common nucleus proper vector of forming; The Classification and Identification that is used for characteristic parameter, this process can represent that wherein a representes easy mistake characteristic parameters such as HighDNAAvgRadius, MedianDNAAvgRadius, LowDNAAvgRadius with Fig. 4.Fig. 5 provides and returns estimation model r among Fig. 4 _jThe acquisition methods of ().

Claims (1)

1. classification of the nucleus under the imperfect situation of characteristic parameter and recognition methods may further comprise the steps: a, concentrated at the nucleus training sample, and distinguishing which characteristic parameter through statistics is to be prone to lose characteristic, which characteristic parameter is the difficult characteristic of losing; It is characterized in that: also comprise

In b, the training process, concentrate, utilize single easy mistake characteristic ginseng value, be difficult for losing characteristic, obtain the recurrence estimation model of single easy mistake characteristic ginseng value through support vector regression SVR training program as input as output at the nucleus training sample;

In c, the identifying, for individual cells nuclear, if the disappearance characteristic parameter is arranged, recurrence estimation model that the utilization training obtains and its not disappearance characteristic parameter return and estimate to be somebody's turn to do the disappearance characteristic parameter;

Specifically: easy mistake characteristic

disappearance of supposition nucleus i calculates estimated value through following method and fills;

(1) extracts this nuclear difficult mistake proper vector b ⁱ

(2) the characteristic parameter a that utilizes training process to obtain _jRecurrence estimation model r _j() returns program through SVR, obtains the estimated value of this characteristic parameter

(3) using the estimated value

fill in the missing characteristic parameter values

Characteristic parameter for these all disappearances of nucleus; All utilize the method for above step (1), (2), (3) to obtain estimated value and fill; After having filled all disappearance characteristic parameters; In conjunction with the easy mistake characteristic parameter that does not lack, obtain being prone to lose the set

of all estimated values of proper vector

D, last; Utilize the difficult mistake proper vector of populated easy mistake proper vector and former extraction, the proper vector

of forming nucleus i jointly is used for nuclear Classification and Identification;

In the above-mentioned steps

I is used for distinguishing different cells nuclear, and j is used for distinguishing the different character parameter,

Represent that i nuclear j is prone to lose characteristic, Represent i nuclear j estimated value that is prone to lose characteristic;

Represent i nuclear all easy set of losing the characteristic parameter estimated value, b ⁱRepresent i nuclear difficult mistake proper vector; a _jThe title of representing j characteristic parameter; f ⁱRepresent the set of i all characteristic parameters of nucleus.

Images